Data & collection
Dataset card
A dataset card is documentation that explains a dataset's contents, collection and processing, intended uses, limitations, and access or licensing terms. It helps users judge whether a particular release fits their task, but does not independently certify its quality or grant rights beyond the applicable licence.
Also known as: dataset cards, data card
Updated
Documentation attached to a release
On the Hugging Face Hub, a dataset card is the repository's README.md, displayed on its main page. YAML metadata can describe the licence, languages, task categories and other properties used for discovery and configuration. The prose supplies context that a file listing cannot: why the data was collected, what it contains, potential biases and appropriate uses.
The related Datasheets for Datasets proposal asks creators to document motivation, composition, collection, processing, uses, distribution and maintenance. A datasheet and a platform's dataset card serve similar documentation goals, but there is no single mandatory schema shared by every repository.
What a robotics buyer needs to learn
For a robotics release, useful card fields include the captured tasks, environments, participants or robot bodies, sensor streams, episode counts, sampling rates and missing signals. The card should explain how labels were produced and link to the file schema, calibration, quality report and data provenance.
For example, a card for egocentric data should distinguish recorded RGB from estimated hand poses and identify whether robot commands exist. An action column containing a human wrist trajectory has a different meaning from a command sent to a robot controller. The card should also describe training and evaluation splits, known tracking failures, privacy processing and permitted uses.
Keep the card tied to a named dataset version. If an export changes the camera model, removes audio or revises labels, its documentation must describe the delivered release rather than an earlier collection plan.
A card supports evaluation, not certification
A well-written card makes claims inspectable; it does not establish that those claims are true. Buyers still need representative files, a working loader and evidence for important quality claims. Likewise, a licence tag summarizes a declared licence. The actual licence or commercial agreement determines the grant, restrictions and obligations. A dataset card cannot establish ownership, participant permissions or downstream redistribution rights by itself.
Sources
Related terms
Data & collection
Data provenance
Data provenance is information about a dataset's origins, the activities that created or transformed it, and the people or systems responsible. It connects delivered files and labels to their source records so users can assess reliability and trace changes; it does not by itself prove ownership or permission to sell the data.
Data & collection
Robot training data
Robot training data is recorded experience used to train, fine-tune, or adapt models for robot perception, prediction, planning, or control. It can include sensor observations, robot state, actions, task instructions, rewards or outcomes, demonstrations, failures, and embodiment metadata. Not every dataset contains every field, but their timing and physical meaning must be clear.
Data & collection
Egocentric data
Egocentric data is sensor data recorded from the viewpoint of the person or robot performing an activity, most commonly with a head- or body-mounted camera. It can also include audio, gaze, depth or inertial signals. For humanoid learning, it shows hands, objects and actions from an actor-centred perspective.
Data & collection
LeRobot dataset
A LeRobot dataset is a robot-learning dataset organised for the LeRobot data model and loaders. LeRobot v3 stores low-dimensional state, action and timestamp fields in Parquet, camera streams in MP4, and schema, task, statistics and episode metadata under a coordinated directory layout.